Prime Intellect raised a $130 million Series A at a $1 billion valuation to sell enterprises the tools they need to train their own AI agents without renting from OpenAI or Anthropic. Radical Ventures led the round, with Nvidia Ventures, Intel Capital, Dell Technologies Capital, and Iconiq participating alongside a roster of founder angels that includes Aravind Srinivas of Perplexity, Aaron Levie of Box, Winston Weinberg of Harvey, Jeff Wang of Cognition, and Brendan Foody of Mercor. The startup is already running at a $100 million annualized revenue rate less than two years after being founded in 2024.
The pitch is that companies no longer need a frontier lab to build capable agentic systems. Reinforcement learning techniques, which reward successful task completion and penalize errors, now let a mid-sized enterprise refine open models for specific workflows. Prime Intellect packages the compute, the RL framework, and the evaluation tooling into what it calls a full stack, sold as modular components rather than an all-or-nothing platform.
Radical Ventures partner David Katz frames the moat as integration rather than any single component. Rival vendors offer compute, or an RL library, or an eval harness — Prime Intellect sells the combination as a marketplace, which is what makes it usable for teams that do not have a research org of their own.
Key facts
- 01Prime Intellect raised a $130M Series A at a $1B valuation, led by Radical Ventures.
- 02Nvidia Ventures, Intel Capital, Dell Technologies Capital, and Iconiq joined the round.
- 03The startup hit a $100M annualized revenue run rate less than two years after its 2024 founding.
- 04Ramp, Zapier, and Flapping Airplanes are paying customers of the hosted platform.
- 05Angel investors include Aravind Srinivas (Perplexity), Aaron Levie (Box), and Winston Weinberg (Harvey).
The customer list is the evidence. Ramp, Zapier, and Flapping Airplanes are paying for a hosted version of the tooling, and Ramp used the platform to build an internal agent that answers questions inside spreadsheets — a workload where frontier models tend to be both slow and expensive.
Ramp co-founder and co-CEO Karim Atiyeh said the resulting agent outperformed frontier models on accuracy while running faster and cheaper. That is the specific promise driving the growth: for a narrow enterprise task, a purpose-trained smaller model beats a general-purpose giant, and the economics do not require an eight-figure inference bill.
“The result beat the frontier models on accuracy while running at faster speeds and a fraction of the cost.”— Karim Atiyeh, Ramp co-founder and co-CEO
The second tailwind is defensive. Enterprises are increasingly nervous about handing proprietary data to a frontier lab that could eventually build a competing product, and about depending on a hosted model that can be deprecated on the vendor's schedule. Anthropic's Fable, which was shut down last month, is the most recent example cited by investors evaluating the space.
Katz put the concern bluntly: customers are asking how they can be sure the lab they build on top of will not eventually generalize to their own use case and compete with them. Owning the weights, or at least owning the training pipeline, is now a board-level question at companies that a year ago were happy to route everything through an API.
Vincent Weisser, Prime Intellect's co-founder and CEO, argues the ability to train frontier-grade models should not be restricted to a handful of labs in San Francisco. His company is betting that a large share of enterprises, and eventually governments, will want to run that pipeline in-house.
The counterweight is that most of Prime Intellect's customers today are technically sophisticated — fintechs and developer-tool companies with strong ML teams already. Whether the same stack works for a regional bank or a manufacturer without a research org is unproven, and the frontier labs are aggressively pricing down inference to keep those customers from ever needing to train their own model in the first place.
Prime Intellect is the clearest venture bet yet on what investors are calling AI sovereignty — the thesis that the next wave of enterprise AI spend goes not to frontier API bills but to the picks and shovels of custom training. The $100 million revenue run rate suggests the demand is already there. The question for the next round is whether that number reflects a durable shift in how enterprises buy AI, or an early cohort of ML-heavy customers who were always going to build in-house.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
